Commonsense Reasoning on CSQA OOD (test)
82.1AccuracyMultiRole-R1
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MultiRole-R1Backbone=Qwen3-8B, Method Strategy=SelfConsis SFT+GRPO(RS)2025.07 | 82.1 | 87.93 | |
| MultiRole-R1Backbone=R1-Distill-Qwen-14B, Method Strategy=SelfConsis SFT+GRPO(RS)2025.07 | 82 | 91.61 | |
| SelfConsis SFTBackbone=R1-Distill-Qwen-14B2025.07 | 81.5 | 77.6 | |
| More thinkBackbone=Qwen3-8B2025.07 | 81.2 | 73.07 | |
| SelfConsis SFT+GRPOBackbone=R1-Distill-Qwen-14B2025.07 | 81.19 | 89.64 | |
| SelfConsis SFT+GRPOBackbone=Qwen3-8B2025.07 | 81.19 | 86.4 | |
| SelfConsis SFTBackbone=Qwen3-8B2025.07 | 81 | 73.5 | |
| Zero-shot CoTBackbone=Qwen3-8B2025.07 | 80.81 | 57.84 | |
| SelfConsis SFT+DPOBackbone=Qwen3-8B2025.07 | 80.45 | 64.87 | |
| SelfConsis SFT+DPOBackbone=R1-Distill-Qwen-14B2025.07 | 79.82 | 69.41 | |
| More thinkBackbone=R1-Distill-Qwen-14B2025.07 | 79.36 | 81.77 | |
| Self-RefineBackbone=R1-Distill-Qwen-14B2025.07 | 76.55 | 82.31 | |
| Zero-shot CoTBackbone=R1-Distill-Qwen-14B2025.07 | 75.85 | 83.09 | |
| Role-PlayBackbone=R1-Distill-Qwen-14B2025.07 | 75.71 | 79.05 | |
| MultiRole-R1Backbone=R1-Distill-Llama-8B, Method Strategy=SelfConsis SFT+GRPO(RS)2025.07 | 75.12 | 92.98 | |
| SelfConsis SFT+GRPOBackbone=R1-Distill-Llama-8B2025.07 | 73.71 | 87.96 | |
| Role-PlayBackbone=Qwen3-8B2025.07 | 72.98 | 48.17 | |
| SelfConsis SFT+DPOBackbone=R1-Distill-Llama-8B2025.07 | 71.28 | 82.71 | |
| SelfConsis SFTBackbone=R1-Distill-Llama-8B2025.07 | 70.86 | 83.88 | |
| More thinkBackbone=R1-Distill-Llama-8B2025.07 | 70.42 | 84.73 | |
| SelfConsis SFT+GRPOBackbone=R1-Distill-Qwen-7B2025.07 | 69.43 | 86.85 | |
| Self-RefineBackbone=Qwen3-8B2025.07 | 69.22 | 50.99 | |
| Role-PlayBackbone=R1-Distill-Llama-8B2025.07 | 67.32 | 82.27 | |
| SelfConsis SFT+DPOBackbone=R1-Distill-Qwen-7B2025.07 | 67.24 | 69.83 | |
| Zero-shot CoTBackbone=R1-Distill-Llama-8B2025.07 | 67.21 | 83.39 | |
| MultiRole-R1Backbone=R1-Distill-Qwen-7B, Method Strategy=SelfConsis SFT+GRPO(RS)2025.07 | 66.94 | 87.96 | |
| SelfConsis SFTBackbone=R1-Distill-Qwen-7B2025.07 | 66.88 | 83.1 | |
| More thinkBackbone=R1-Distill-Qwen-7B2025.07 | 64.5 | 85.85 | |
| Self-RefineBackbone=R1-Distill-Llama-8B2025.07 | 63.77 | 82.61 | |
| Zero-shot CoTBackbone=R1-Distill-Qwen-7B2025.07 | 63.06 | 83.83 | |
| Role-PlayBackbone=R1-Distill-Qwen-7B2025.07 | 55.07 | 76.2 | |
| Self-RefineBackbone=R1-Distill-Qwen-7B2025.07 | 54.02 | 77.61 |